You may have already figured this out but your assumption is right - Nearest Neighbor would be the best resampling method for geometric correction. This way, you would not be affecting the original values, which have already been affected by the original resampling performed USGS. On a similar note, Nearest Neighbors would be the only approach you would want to use for any classifications or similar discrete value rasters for the same reason as above. Otherwise, your neat, integer class numbers (0, 1, 2, 3, 4) would be lost through interpolation.
If your time series is derived entirely from USGS and they have all undergone geometric corrections, there may be no need to perform additional co-registration - most users have found that USGS georeferencing is quite robust.